> ## Content Index
> Fetch the complete content index at: https://www.antoinebuteau.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Daily Digest - 2026-06-19
- URL: https://www.antoinebuteau.com/daily-digest-2026-06-19/
- Published: 2026-06-20T11:00:51.000Z
- Updated: 2026-08-13T01:01:30.000Z
- Description: AI economics increasingly favor engineered workflows over premium models alone. Specialized skills, automated loops, and explicit inference budgets can make capable systems cheaper, provided teams measure productivity against cost and design the surrounding machinery carefully.
- Author: Antoine Buteau
- Tags: Digest

**1\. How I turned MiniMax into Fable 5 (97% cheaper) — anita · vellum.ai 👾🦾**

- Why read: See how you can get top-tier AI performance on a budget by pairing open models with specific skills and plugins.
- Summary: The author created a 97% cheaper alternative to Fable 5 using the MiniMax model. The trick is giving the model specific context and tools via well-written SKILL.md files and plugins. MiniMax handles tool use and complex tasks like building CLI tools or turning call transcripts into agents. It isn't great at writing by default, but teaching it specific writing skills fixes that. The main takeaway: you can cut inference costs significantly by building specialized workflows around less expensive models.
- [Read more](https://twitter.com/anitakirkovska/status/2067654220027744623/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

**2\. From Prompting Agents to Loop Engineering — elvis**

- Why read: Understand the move from writing manual prompts to building systems that prompt the models automatically.
- Summary: AI coding is shifting from manual prompting to building "loops": small programs that set a goal, act, check the output, and feed errors back to the model until the job is done. Developers are becoming system orchestrators instead of direct operators. A functional loop needs triggers, isolated workspaces, and written context to guide the agent. This setup boosts productivity and lets developers merge multiple PRs without writing much code by hand.
- [Read more](https://twitter.com/omarsar0/status/2068008743153832264/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

**3\. You should ask for an inference budget — Greg & Taylor**

- Why read: See why teams need a dedicated "inference budget" to manage AI costs and measure returns.
- Summary: Managers need a specific budget for AI inference, treated like a fraction of an employee's salary. A starting point is $250 per employee, rising to $1,000 or $2,000 for heavy users like engineers. The aim is to get more productivity out of the AI than you would from hiring another person. Instead of worrying about every token, teams should focus on connecting top models to company data. This treats AI as a standard business expense with measurable results.
- [Read more](https://read.readwise.io/read/01kvgcnb6fv97gv4t5h58613b0?ref=antoinebuteau.com)

**4\. 644: Apple + Intel, OpenAI's IPO Financials, Tim Cook's RAM Headache, Benedict Evans Interview, Snap AR Glasses, W… — Liberty’s Highlights**

- Why read: How the culinary ideas of "solera" and "perpetual stew" explain long-lasting tech systems.
- Summary: A "perpetual stew" is never fully emptied; new ingredients are just added to the old. The author uses this as a metaphor for tech systems that evolve but keep traces of their past. This applies to older components left in modern hardware and continuous software updates. Understanding these legacy elements shows how they shape current versions. Recognizing what persists helps teams build more adaptable systems.
- [Read more](https://read.readwise.io/read/01kvgdts3h029tz8vxf3ntc71k?ref=antoinebuteau.com)

**5\. \[AINews\] GLM > GPT GLM-5.2 passes vibe check; Z.ai forecasts Open Fable by December — AINews**

- Why read: Track how new open-source models are stacking up against proprietary options.
- Summary: GLM-5.2 is passing the "vibe check" with practitioners. Many open models ace benchmarks but fail in real use, but GLM-5.2 is actually outperforming GPT-5.5 on knowledge work. This means an open model is finally competing directly with top-tier closed systems. On top of this, forecasters predict an open-source match for Fable by December. Teams should start testing these models now to cut costs and avoid vendor lock-in.
- [Read more](https://read.readwise.io/read/01kvf73wkb0befmce6zt4qbez4?ref=antoinebuteau.com)

**6\. Fully Automated Prompt Optimization — Cisco Foundation AI**

- Why read: How to stop tweaking prompts by hand and automate the process.
- Summary: Nailing the right prompt is hard; tiny edits can tank a model's accuracy. Manual tweaking doesn't scale and creates fragile systems. FAPO (Fully Automated Prompt Optimization) uses Claude Code agents to evaluate failures and rewrite prompts automatically. You supply a test dataset and a starting prompt, and the system loops until it hits your accuracy target. This generates stable prompts without the manual trial and error.
- [Read more](https://cisco-foundation-ai.github.io/blogs/fully-automated-prompt-optimization/?ref=antoinebuteau.com)

**7\. The Stuff of Myth(os) (This Week in Stratechery) — Ben Thompson**

- Why read: A look at Anthropic's market position and how AI is changing e-commerce.
- Summary: Export controls recently forced Anthropic to limit access to its Fable model, showing how geopolitics affect AI. Anthropic's focus on safety is both a genuine mission and a strategic advantage that frames its business moves as principled. The article also covers how AI is shifting traditional e-commerce moats. Companies need to understand these dynamics, especially how safety and alignment are now used as competitive weapons.
- [Read more](https://stratechery.com/2026/the-stuff-of-mythos/?ref=antoinebuteau.com)

**8\. The Certifiably Insane Way to Build an AI Agent: — austin petersmith**

- Why read: How one team brute-forced edge cases with human reviewers to build a reliable AI agent.
- Summary: To build a dependable AI scheduling assistant, the team kept humans deeply involved to handle edge cases. In scheduling, mistakes cost users prospects or investors, so trust is everything. They hired contractors to catch errors and map out resolutions for millions of specific scenarios. This heavy lifting—collecting data, fine-tuning, and iterating—resulted in a system that now runs 50% on autopilot with high quality. The takeaway: using human oversight to grind out a zero-defect product builds a massive data advantage.
- [Read more](https://twitter.com/awwstn/status/2067750610963939571/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

**9\. Clouded Judgement 6.19.26 - Workflows are King — Clouded Judgement by Jamin Ball**

- Why read: Why owning the workflow is now more defensible than owning the data.
- Summary: SaaS companies used to build moats by being the primary data storage system. But what actually kept users around were the workflows built on top of that data. With AI agents, the moat is moving from storing data to orchestrating work. Founders shouldn't try to build a massive platform right away. Instead, they need to master one highly specific workflow, claim that ground, and slowly expand to adjacent tasks.
- [Read more](https://read.readwise.io/read/01kvg0fqcpepg33c3wy6jxebg8?ref=antoinebuteau.com)

**10\. $400M raised. 750 jobs cut. And the stat that explains why most AI pivots fail. — Sangram**

- Why read: Why adding AI to a bad Go-To-Market strategy just makes things fail faster.
- Summary: Companies are pouring money into AI while cutting jobs, often misunderstanding how AI fits into Go-To-Market (GTM) strategy. Slapping AI onto a broken GTM system doesn't fix it; it just speeds up the mistakes. Most companies struggle because their basic operations are flawed, not because they lack AI. You have to fix the foundation and build a clear operating system first. Once things work efficiently, you can use AI to speed them up.
- [Read more](https://read.readwise.io/read/01kvfwbztxdj31j4zas9ey4xqk?ref=antoinebuteau.com)

**11\. Craft Irresistible Outbound Campaigns Using Claude Code — Maja Voje from GTM Strategist**

- Why read: How analyzing user conversations with AI can fix failing outbound campaigns.
- Summary: Outbound campaigns usually fail because they send generic messages to the wrong people. But enterprise buyers actually reveal much more to AI agents than they do to human salespeople. How a buyer talks to an AI can predict their seniority and the deal's size. Market demand is also highly concentrated, and conversational data captures this intent far better than static web forms. Using these chat signals lets teams build highly targeted outbound campaigns.
- [Read more](https://read.readwise.io/read/01kvfj7fj2c001n8dryhs6tfv3?ref=antoinebuteau.com)

**12\. Inspect at Scale — Dylan Garcia**

- Why read: How an internal coding agent grew to handle 76.5% of all PRs at Ramp.
- Summary: Ramp's coding agent, Inspect, now touches over three-quarters of the company's merged PRs. It's used by engineering, product, design, and data teams for everything from reviewing code to building web apps. The team analyzed 800,000 sessions to find where the agent fails and improve its reliability. By obsessing over user experience details like sandboxes and cold starts, they turned a narrow utility into a platform the whole company uses.
- [Read more](https://twitter.com/%5Fdylanga/status/2067999346780524753/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

**13\. Building Autonomous Agent Loops — elvis**

- Why read: The exact pieces you need to build an autonomous agent that actually works.
- Summary: You can't just prompt a model; you need to build an agent loop out of reliable parts. A real loop needs an automated trigger like a webhook or a cron schedule. It also needs isolation—agents should work in private environments like git worktrees so they don't break things. Finally, you have to write down the context and rules so the agent follows project conventions. Getting these parts right lets you build systems that handle long tasks without a human watching them.
- [Read more](https://twitter.com/omarsar0/status/2068008743153832264/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

**14\. The Agentic Services Company Paradigm — Sangram**

- Why read: Why the most successful Go-To-Market teams will combine human services with AI agents.
- Summary: A "human-agentic services company" recently raised $400M, signaling a new way to execute work. Many companies fail at AI because they think the tools alone will fix their execution problems—like firing staff just to plug in a bot. But execution is getting harder as timelines shrink and channels multiply. The correct approach is fixing the organization's foundation first. AI only provides a real edge when deployed as an execution layer on top of a system that already works.
- [Read more](https://read.readwise.io/read/01kvfwbztxdj31j4zas9ey4xqk?ref=antoinebuteau.com)

**15\. The Importance of Written-Down Context for AI — elvis**

- Why read: Why forcing agents to read explicit documentation is the secret to making them work in production.
- Summary: Agents don't have the implicit knowledge human developers do about build steps, conventions, or past decisions. If you don't write this context down for the agent to read constantly, it will get lost in its own history. Managing this context explicitly keeps agents focused and accurate on long tasks. You have to treat the agent like a brand-new employee: give it clear, accessible documentation if you want it to succeed.
- [Read more](https://twitter.com/omarsar0/status/2068008743153832264/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

### Themes from yesterday

- **From Prompting to Orchestration:** We are moving from writing manual prompts to building automated loops that manage the models for us.
- **Humans for the Edge Cases:** In high-stakes situations, using humans to map out and handle edge cases works better than expecting AI to be perfect immediately.
- **AI Breaks Bad Systems Faster:** Adding AI to a flawed process just speeds up the mistakes; you have to fix the core operations before layering AI on top.